how to choice the index type when create collection?

[hxlsky]

how to choice the index type when create index,which factor need to consider?

[yhmo]

• IVF_FLAT, memory usage is a bit larger than the original data size, "nprobe" value affects recall rate and search speed
• HNSW, memory usage is a bit larger than the original data size, high recall rate, high speed
• IVF_SQ8/IVF_PQ, memory usage is 25% ~ 30% of the original data size, the recall rate of IVF_PQ is not good.
• DISKANN, memory usage is 1/4 ~ 1/6 of the original data size, performance requires high-speed hard disk.

You can estimate the original data size in this way: each dimension is represented as a float32 value, which occupies 4 bytes. So, the original data size is: row_count * dimension * 4 bytes